Testing the Mother's Curse Hypothesis in Human Mitochondrial Genome Evolution

Abstract

In species where mitochondrial DNA (mtDNA) is maternally inherited such as vertebrates, mtDNA mutations harming males only are not subject to purifying selection and thus can spread in a population, especially when these mutations benefit females. Therefore, the mother's curse hypothesis (MCH) posits a greater mtDNA mutation load in males than in females. MCH is potentially important for human health, disease, and evolution, but a systematic test that considers the vast human mtDNA variation is lacking. Analyzing the genotypic and phenotypic data of approximately 0.5 million British participants in the UK Biobank, we estimate the reproductive fitness of mtDNA variants in each sex. Contradicting MCH, a positive intersexual correlation in the number of offspring exists across mitochondrial haplogroups. While a significant variation in the number of opposite-sex sexual partners—a proxy for reproductive fitness in premodern societies—is present among mitochondrial haplogroups, no significant intersexual correlation in this quantity is detected. The frequencies of a few mtDNA variants differ significantly between males and females, suggesting that these variants differentially affect the survival in the two sexes, but the number of such variants with lower male frequencies is not significantly different from that with lower female frequencies. Analysis of disease associations also finds no enrichment of male disease-associated mtDNA variants despite the discovery of multiple sex-biased disease associations. Together, these findings provide no genomic support to MCH in humans and suggest no difference in mtDNA mutation load between the two sexes that is detectable in the UK Biobank.

Significance.

The mother's curse hypothesis (MCH) posits a higher mitochondrial DNA (mtDNA) mutation load in males than in females, because mtDNA mutations deleterious to males only are invisible to natural selection in species where mtDNA is maternally transmitted. Here, we perform a comprehensive genomic test of MCH in humans using genotypic and phenotypic data of about 0.5 million participants in the UK Biobank. Our analyses of the reproductive effects, survival effects, and disease associations of mtDNA variants show patterns inconsistent with predictions of MCH and suggest that MCH is unsupported by the UK Biobank data.

Introduction

Mitochondrial DNA (mtDNA) is inherited through the maternal lineage in most multicellular organisms including humans (Giles et al. 1980; Pagnamenta et al. 2021; Sasaki and Sato 2021). The mother's curse hypothesis (MCH) states that, in species where mtDNA is maternally inherited, mtDNA mutations that lower male fitness but not female fitness are not subject to purifying selection so can spread in a population and accumulate in evolution, leading to a greater mtDNA mutation load in males than in females (Frank and Hurst 1996; Gemmell et al. 2004). There are two forms of MCH (Havird et al. 2019). The strong form involves mtDNA mutations beneficial to females but deleterious to males, while the weak form involves mutations neutral to females but deleterious to males. Hereinafter, we refer to mtDNA variants deleterious to males but neutral or beneficial to females as MC variants, which are further divided into strong (beneficial to females) and weak (neutral to females) MC variants. Note that three terms are commonly used to describe mtDNA variants: mitochondrial single-nucleotide variants (mtSNVs), haplotypes, and haplogroups. An mtSNV is a single-nucleotide polymorphism in the mitochondrial genome. A mitochondrial haplotype is a unique mtDNA sequence, whereas a mitochondrial haplogroup is a group of mitochondrial haplotypes sharing a particular set of mtDNA variants (Torroni et al. 2020). Grouping multiple mitochondrial haplotypes into a haplogroup increases the group size and hence sample size, but could sometimes miss important differences among haplotypes of the same haplogroup.

The strongest evidence for MCH has been obtained from the fruit fly Drosophila melanogaster (Innocenti et al. 2011; Patel et al. 2016; Camus and Dowling 2018; Carnegie et al. 2021). For instance, Patel et al. (2016) reported a weak MC variant, although they provided no information on the general prevalence of MC variants. Camus and Dowling detected a negative intersexual genetic correlation in fitness across 13 mitochondrial haplotypes that were placed in the same nuclear background (Camus and Dowling 2018). While this observation is consistent with MCH, it does not reveal the sex suffering from a higher mtDNA mutation load. More recently, Carnegie et al. (2021) reported that, across multiple nuclear genetic backgrounds, mitochondrial genetic variance for wing size is larger in male than female flies. While this observation is compatible with the notion that MC variants are prevalent in flies, it does not provide direct evidence for MCH.

However, there is also evidence against MCH (Immonen et al. 2016; Mossman et al. 2016a, 2016b; Đorđević et al. 2017; Watson et al. 2022). For example, a study in the copepod Tigriopus californicus reported a higher mitochondrial genetic variance for fertility and longevity in females than in males, contradicting the prediction of MCH (Watson et al. 2022). Further, there was little evidence of sexual antagonism favoring females, and mitonuclear mismatch harmed females but not males (Watson et al. 2022). Theoretically, it has been shown that inbreeding and kin selection can reverse mother's curse (Unckless and Herren 2009; Wade and Brandvain 2009). Hence, the importance of MCH likely varies across different species and populations.

MCH, if true in humans, would have important implications for human health, disease, and evolution, but human studies on the topic have produced mixed results. One case study reported that human mtDNA mutation from T to C at position 14484 (T14484C) is a weak MC variant and that the frequency of the C allele has risen slightly (albeit nonsignificantly) in the last three centuries (Milot et al. 2017). However, this study focused on only one variant and was based on a relatively small sample (97 carriers). There have been several reports that certain human mitochondrial haplogroups are associated with male infertility. For instance, haplogroups H and T are overrepresented in nonasthenozoospermic (i.e. normal sperm motility) and asthenozoospermic (i.e. reduced sperm motility) samples, respectively (Ruiz-Pesini et al. 2000), and haplogroup U5b shows significantly higher sperm motility than haplogroup H (Montiel-Sosa et al. 2006). However, there is also evidence that haplogroups are not associated with sperm motility (Mossman et al. 2012). A major caveat of these association studies is that false positives may result if population structure is not appropriately controlled (Pereira et al. 2005).

The advent of large biobanks with rich genotype and phenotype information of humans has offered an unprecedented opportunity to perform a systematic test of MCH that considers the vast human mtDNA variation, which could shed light not only on MCH but also on potential mtDNA mutations in male diseases, including male infertility (Ruiz-Pesini et al. 2000). Mitochondrial genome-wide association study (mtGWAS) and haplogroup-based association analysis are two potential approaches to testing MCH in humans. mtGWAS examines the association between mtSNVs and phenotypes, while haplogroup-based association analysis tests the association between haplogroups and phenotypes. mtGWAS can be more powerful than haplogroup-based association analysis because there are fewer mtSNVs than haplogroups. However, results from haplogroup-based association analysis are more interpretable because the entire mitochondrial genome is linked (Eyre-Walker and Awadalla 2001). While a previous mtGWAS using the UK Biobank (UKB) (Bycroft et al. 2018) found no mtDNA variant to be significantly associated with the number of offspring (Yonova-Doing et al. 2021), the authors lumped the two sexes in the analysis, so they could have missed MC variants. Furthermore, it is unknown whether mtDNA variants are associated with other fitness components such as survival. In the present study, we systematically test MCH in humans using the UKB. Our results provide no genomic support to MCH, suggesting that, to the extent that is discernible from the UK Biobank, mtDNA mutation load is similar in the two sexes.

Results

Association Between Mitochondrial Haplogroups and the Number of Offspring

The UKB contains the nuclear and mitochondrial genotypic data and various phenotypic data of about 0.5 million participants from the United Kingdom (Bycroft et al. 2018). Using the mitochondrial genotyping data, we classified the mtDNA of an individual into a haplogroup; 378,777 individuals, including 204,240 females and 174,537 males, passed our quality cutoff (see Materials and Methods).

The strong form of MCH predicts an enrichment of mtDNA variants with sexually antagonistic fitness effects, while the weak form predicts an overrepresentation of mtDNA variants that harm males only. We first considered reproductive fitness measured by the number of offspring. Note that a person's number of offspring is this person's realized fitness, although it may differ from this person's fertility due to many factors including socioeconomic factors. For each sex, we built a linear model to predict the number of offspring of an individual from this person's top ten nuclear genetic principal components (PCs), year of birth, and mitochondrial haplogroup (see Model 1 in Materials and Methods). A total of 603 haplogroups had at least one sample in each sex and were included in the analysis. Haplogroup H1 was chosen to be the reference haplogroup because (i) H1 is the haplogroup with the largest sample size and (ii) the mean number of offspring of individuals of H1 is not significantly different from that of all other haplogroups combined in each sex (P = 0.56 for males and 0.70 for females, two-tailed t-test; Table S1). The regression coefficient of a haplogroup, referred to as haplogroup effect on offspring number (HEON), is the effect of the haplogroup on the number of offspring minus that of haplogroup H1 in the sex concerned. Fifty-two haplogroups exhibit nonzero HEON in at least one sex at nominal P < 0.05 (Fig. 1a), but none of them are significant at the false discovery rate (FDR) of 0.05 after correction for multiple testing (602 tests) by the Benjamini–Hochberg procedure (Benjamini and Hochberg 1995), suggesting that no single haplogroup is significantly associated with the number of offspring in either sex. We additionally performed omnibus tests (Type II ANOVA tests) to evaluate the overall effect of mitochondrial haplogroups on the number of offspring. We found no evidence against the null hypothesis that all haplogroups have the same effects on the number of offspring when controlling the other predictor variables mentioned earlier (P = 0.51 for males and 0.67 for females).

Fig. 1.

HEON relative to the effect of haplogroup H1. a) Haplogroups and their effects in each sex. Shown are the 52 haplogroups (ordered by female HEON) with effects different from 0 at nominal P < 0.05 in at least one sex. Error bars represent standard errors. Note that no HEON is significant after FDR correction. U5a2 + 16362 indicates an unnamed haplogroup, which is a variant of haplogroup U5a2 harboring an additional transition at mtDNA site 16362. H + 13708 indicates an unnamed haplogroup, which is a variant of haplogroup H harboring an additional transition at mtDNA site 13708. b) HEON values of all 602 haplogroups. Note that dots show HEON values without phylogenetic corrections. The solid line and the annotation show the best-fit under PGLS regression. The shaded area shows the 95% confidence interval of the regression. Note that the positive intersexual correlation would become nonsignificant if the seven HEONs with the highest (top 1%) leverage are excluded from the PGLS regression.

To test whether any individual haplogroup has a sex-differential effect on the number of offspring, we built an interaction model including sex × haplogroup terms (see Model 2 in Materials and Methods). The model predicts the number of offspring of an individual by this person's top ten genetic PCs, year of birth, mitochondrial haplogroup, sex, and the interaction between sex and each of the other predictor variables. Consistent with the results of the sex-specific models, no sex × haplogroup interaction term remains significant at an FDR of 0.05 (602 tests), indicating that no haplogroup has a significant, sex-differential effect on the number of offspring.

We measured the intersexual correlation in HEON across all haplogroups, which yielded a nonsignificant result (Spearman's ρ = −0.015, P = 0.71). However, because the haplogroups studied are evolutionarily related, we tested whether the HEON values show a phylogenetic structure. Specifically, we first predicted female HEON from male HEON with an ordinary least squares (OLS) model and then estimated Pagel's λ (Pagel 1999) for the residuals of the model given the phylogenetic tree of the haplogroups (Fig. S1). A λ of 0 means that trait covariances among tips in the tree are independent of phylogenetic relationships, whereas λ = 1 means that the covariances are predicted exactly by shared ancestry under a Brownian motion model. There was indeed a significant phylogenetic signal in the residuals (λ = 0.93, P = 5.0 × 10⁻¹⁵). We therefore tested the intersexual correlation in HEON using phylogenetic generalized least squares (PGLS), which controls the phylogenetic relationships among the haplogroups. The PGLS assumes a Brownian motion model for evolutionary changes of the trait along tree branches and predicts the female HEON from the male HEON of each haplogroup. Contrary to the expectation of a negative intersexual correlation under MCH, the regression coefficient is significantly positive (P = 0.0003; Fig. 1b), suggesting that intersexual concordance rather than antagonism is the predominant pattern of mitochondrial haplogroups’ effects on the number of offspring.

Association Between Mitochondrial Haplogroups and the Number of Sexual Partners

Although the number of offspring recorded in the UKB measures the present reproductive fitness, it probably poorly reflects the reproductive fitness in premodern societies, due to the widespread use of contraception in the United Kingdom for the past 60 years or so (Song and Zhang 2023). Because the present-day human mtDNA variation is a result of evolution over a long time, it is important to also estimate premodern reproductive fitness when testing MCH. It has been suggested that the number of opposite-sex sexual partners (number of sexual partners hereinafter) can be used as an approximate measure of the reproductive fitness in premodern societies (Zietsch et al. 2021), because under polygyny, which was common in human history, a male with more sexual partners tends to have more offspring (Gibson and Mace 2007; von Rueden and Jaeggi 2016). The number of sexual partners of a female also positively predicts her number of offspring in traditional human societies (Scelza 2011).

Using the method described in the preceding section, we estimated the effect of each haplogroup on the number of sexual partners minus that of haplogroup H1 in the sex concerned; this effect is referred to as the haplogroup effect on partner number (HEPN). We again used haplogroup H1 as the reference because (i) H1 is the haplogroup with the largest sample size and (ii) the mean number of sexual partners of individuals of H1 is not significantly different from that of all other haplogroups combined in each sex (P = 0.11 for males and 0.87 for females, two-tailed t-test; Table S2). We included in the analysis all 581 haplogroups that each had at least one sample per sex. Sixty-six haplogroups exhibit nonzero HEPN in at least one sex at nominal P < 0.05 (Fig. 2a), but none of them have significant HEPN in either sex with adjusted P < 0.05 after the FDR correction for 580 tests. Haplogroup H2a2b5a has a marginally significant HEPN in males (adjusted P = 0.051) and a nonsignificant HEPN in females (nominal P = 0.077), but its effect is positive in males and negative in females, contrary to the expectation for an MC variant (Fig. 2a). We additionally performed omnibus tests (Type II ANOVA tests) to evaluate the overall effect of mitochondrial haplogroups on the number of sexual partners. Our results suggest that at least one haplogroup has significantly different effects on the number of sexual partners in the two sexes (P = 0.013 for males and 0.009 for females). Hence, mtDNA variation might affect reproductive fitness in ancient times. We also built a similar interaction model as described in the preceding section to detect sex × haplogroup interaction effects on the number of sexual partners. Again, we found the sex × H2a2b5a interaction to be the only significant sex × haplogroup interaction term (adjusted P = 0.007) after an FDR correction for 580 tests. However, the interaction effect of H2a2b5a is opposite to the expectation for an MC variant.

Fig. 2.

Haplogroup effects on opposite-sex sexual partner number (HEPN) relative to the effect of haplogroup H1. a) Haplogroups and their effects in each sex. Shown are the 66 haplogroups (ordered by female HEPN) with effects different from 0 at nominal P < 0.05 in at least one sex. Error bars represent standard errors. Note that no HEPN is significant after FDR correction. T2 + 150 indicates an unnamed haplogroup, which is a variant of haplogroup T2 harboring an additional transition at mtDNA site 150. b) HEPN values of all 580 haplogroups. Note that dots show HEPN values without phylogenetic corrections. The solid line and the annotation show the best-fit under PGLS regression. The shaded area shows the 95% confidence interval of the regression.

We further measured the intersexual correlation in HEPN across all haplogroups, which yielded a nonsignificant result (Spearman's ρ = 0.026, P = 0.53). Because we found a significant phylogenetic signal in the residual of the OLS model predicting female HEPN from male HEPN (λ = 0.97, P = 1.2 × 10⁻¹⁸), we again built a PGLS model to control phylogenetic relationships. The PGLS yielded a nonsignificant intersexual correlation (P = 0.62; Fig. 2b). In sum, the analysis of the number of sexual partners, an approximate measure for premodern reproductive fitness, lends no support to MCH either.

Association Between mtSNVs and Reproductive Fitness

We then used mtGWAS to test whether individual mtSNVs, instead of mitochondrial haplogroups, are associated with reproductive fitness measured by the number of offspring as well as that measured by the number of sexual partners, in each sex and in the two sexes combined. No single mtSNV showed a significant association (adjusted P < 0.05 after correcting for 265 tests) in any category, providing no support to MCH.

Differential Frequencies of mtDNA Variants in the Two Sexes

We next examined whether mtDNA variation is associated with survival, another important component of fitness. Because all UKB participants had lived to at least 40 years old at the time of their UKB participation, potential differential survival before the age of 40 cannot be directly assessed using the UKB data. Notwithstanding, the mitochondrial genotype of an egg is unlikely to affect whether the egg is fertilized by an X chromosome-carrying sperm or a Y chromosome-carrying sperm. Hence, at conception, males and females are expected to have equal frequencies of any mtDNA variant. If an mtDNA variant affects survival differentially in the two sexes, the variant would have different frequencies in the two sexes among adults. That is, we can assess sex-differential survival associated with an mtDNA variant by sex-differential frequencies of the variant (Cole et al. 2024). Therefore, MC variants that are deleterious to male survival but are neutral or advantageous to female survival should show lower allele frequencies in adult males than in adult females.

We identified one mtSNV and two mitochondrial haplogroups that exhibit significant sex-differential frequencies based on a chi-squared test followed by FDR correction for 414 tests for haplogroups and 241 tests for mtSNVs (Table S3). G3316A has a significantly lower frequency in males (0.20%) than in females (0.27%), consistent with a previous report (Yonova-Doing et al. 2021). Haplogroup U5b1b1a has a significantly lower male frequency, while haplogroup T1 has a significantly higher male frequency (Fig. 3; Table S3). Because the number of mtSNVs or mitochondrial haplogroups supporting MCH (1 mtSNV plus 1 haplogroup) does not significantly exceed that contradicting MCH (1 haplogroup), our analysis of the mtDNA variation impacting survival provides no support to MCH.

Fig. 3.

Haplogroup frequencies in males and females. In total, 414 haplogroups with at least six samples in each sex are included. The diagonal line indicates equality in frequency between the sexes. Two haplogroups with significantly different frequencies in the two sexes (adjusted P < 0.05 after correcting for 414 tests) are labelled. Note that the mtSNV (G3316A) with significantly unequal frequencies in males and females is detailed only in the main text and Table S3.

Sex-specific Associations Between mtDNA Variants and Diseases

The preceding test of MCH is expected to have a limited power, because it works only when the differential selection on survival causes a significant allele frequency difference between the two sexes within a generation (i.e. from fertilized egg to adulthood). This means that the test may miss weakly differential selections. We thus further tested whether there are more mtSNVs associated with diseases in males than in females, as MCH predicts. To this end, we first repeated mtGWAS in the two sexes combined for 19 diseases (Table S4) associated with at least one mtSNV in a previous mtGWAS that did not separate the two sexes (Yonova-Doing et al. 2021). We then performed these mtGWAS for males and females, respectively. We found that seven mtSNVs are significantly associated with five diseases in males and six mtSNVs are significantly associated with five diseases in females (Table 1). This approximate parity between the two sexes provides no support to MCH.

Table 1.

mtSNVs associated with diseases in each sex

Disease-associated mtSNV	Female			Male
	Frequency	Odds ratio	Adjusted P	Frequency	Odds ratio	Adjusted P
Volvulus
A827G	2.0 × 10−3	5.45	0.007	1.9 × 10−3	1.03	1
T6253C	1.9 × 10−3	1.50	1	1.8 × 10−3	6.54	0.03
Abnormal findings on diagnostic imaging of other parts of digestive tract
T10321C	2.9 × 10−4	1.30	1*	2.1 × 10−4	14.86	0.02
Calculus of kidney
T16172C	0.040	2.56	0.03	0.040	1.38	0.80
Urinary tract infection/kidney infection
A12810G	3.6 × 10−3	0.91	1	3.3 × 10−3	4.39	0.03
Multiple sclerosis
A14133G	9.9 × 10−3	2.26	0.03	0.010	1.04	1
A10398G	0.21	1.30	0.05	0.21	0.95	1
Ptosis of eyelid
C15452A	0.21	1.29	0.04	0.21	1.35	0.03
T4216C	0.21	1.26	0.08	0.21	1.34	0.04
A11251G	0.21	1.27	0.08	0.21	1.35	0.04
Descending colon
A12397G	2.3 × 10−3	3.64	0.04	2.1 × 10−3	1.31	1
Bladder problem (not cancer)
T7175C	5.4 × 10−4	1.35	1	5.2 × 10−4	10.53	0.04

mtSNV, mitochondrial single-nucleotide variant (e.g. A827G: A to G mutation at mtDNA site 827). Frequency, alternative allele frequency. Adjusted P-values below 0.05 are bolded. * indicates that logistic regression fails to converge so Firth regression is used.

Of the 12 mtSNVs significantly associated with a disease in at least one sex, only one mtSNV (C15452A) is significantly associated with a disease in both sexes (Table 1). Two additional mtSNVs (T4216C and A11251G) are marginally associated with the disease in the other sex (adjusted P = 0.08). Specifically, ptosis of eyelid is associated with three mtSNVs, namely C15452A, T4216C, and A11251G. C15452A and A11251G are diagnostic of haplogroup JT. T4216C is diagnostic of haplogroup R2′JT, which is the direct ancestor of haplogroup JT (van Oven and Kayser 2009). Furthermore, the three mtSNVs have similar frequencies. Thus, the causal mtSNV for ptosis of eyelid is likely associated with the haplogroup R2'JT that shares the three mtSNVs. Hence, other than the association with ptosis of eyelid, mtSNVs affect males and females differently in disease. Nonetheless, the result does not support MCH because mtDNA variants do not cause diseases more in males than in females.

Discussion

We conducted a comprehensive test of MCH using the genotypic and phenotypic data in the UKB. We did not find any mitochondrial haplogroup to be significantly associated with the number of offspring—a measure of the present-day reproductive fitness—and did not find any haplogroup to have a significant, sex-differential effect on the number of offspring. A significant, positive intersexual correlation exists for the haplogroup effect on the number of offspring, contradicting the prediction of MCH. We found no mitochondrial haplogroup to be significantly associated with the number of sexual partners in either sex—an approximate measure of the reproductive fitness in premodern societies. We detected sex-by-haplogroup interaction for one haplogroup (H2a2b5a), but the interaction was opposite to the direction predicted by MCH. No mtSNV was found to be significantly associated with the number of offspring or number of sexual partners in either sex. An analysis of allele frequency differences between the two sexes, an indication that the alleles are associated with differential survivals in the two sexes, did not yield significantly more mtSNVs (1 vs. 0) or mitochondrial haplogroups (1 vs. 1) that have lower frequencies in males than those that have lower frequencies in females. An mtGWAS of diseases performed separately for the two sexes showed that, although mtSNVs tend to be associated with diseases differently in the two sexes, there is no evidence that they cause disease more often in males than in females. Taken together, our results provided no evidence for MCH. Note that our analyses employed mtSNV-based mtGWAS and haplogroup-based association analysis to associate mtDNA variants with phenotypic traits. While the two approaches have their respective pros and cons (see Materials and Methods), they provided overall consistent results.

While it is true that mtDNA mutations that harm males only are not subject to purifying selection because of the maternal inheritance of mtDNA in humans, there are several potential nonmutually exclusive explanations why the mtDNA mutation load appears similar in the two sexes. First, as mentioned, mother's curse may be reversed by inbreeding and kin selection (Unckless and Herren 2009; Wade and Brandvain 2009).

Second, given the high similarity in biology between the two sexes, it is likely that most mtDNA mutations have concordant instead of antagonistic effects in the two sexes. The paucity of mtDNA mutations that harm males but not females may be the primary reason why we found no genomic evidence for MCH. Supporting this view is the observation of a significant positive intersexual correlation of the effect on the number of offspring across mitochondrial haplogroups in the PGLS analysis. Consistently, a nuclear GWAS study in humans also found that magnitude differences, rather than sign differences, are the primary mode of gene-by-sex interaction in complex traits (Zhu et al. 2023).

Third, most humans harbor heteroplasmic mtDNA variants in somatic cells (Gupta et al. 2023). Human eggs are also likely to be heteroplasmic with de novo mutations because the mtDNA mutation rate is 2.87 × 10⁻⁶ mutation/bp/generation (Árnadóttir et al. 2024) and the average mtDNA copy number is 7.95 × 10⁵ (± 2.43 × 10⁵) in metaphase II oocytes (Barritt et al. 2002). Even though we infer only one mitochondrial haplogroup from a person's genotyping data, it remains possible that one or more other mitochondrial haplogroups exist in the person at lower frequencies than that of the inferred haplogroup. Under this scenario of heteroplasmy, if the primary haplogroup conforms to MCH, its deleterious effect in males could be partially masked by normally functioning haplogroups and therefore becomes harder to detect.

Fourth, nuclear compensation may mask MC variants’ deleterious effects on males. To assess this possibility, we tested the interactive effect between nuclear SNVs (nucSNVs) and mtSNVs on the number of offspring, focusing on males. Because there is a huge number of mtSNV–nucSNV pairs with potential epistasis, it is not feasible to test epistatic effects for all of them. To restrict the candidate set of epistatic nucSNVs, we first identified, at genome-wide significance (nominal P < 5 × 10⁻⁸), 12 nucSNVs associated with the number of offspring in males. Because nucSNVs with epistatic effects tend to have additive effects (Bloom et al. 2015), we built interaction models with these 12 nucSNVs and all 265 mtSNVs to predict the number of offspring in each sex (see Materials and Methods). The interaction terms in the models indicate nucSNV–mtSNV interaction, but no interaction term is significantly different from 0 after the FDR correction for 2,859 tests. Using the same method, we built similar interaction models to predict the number of sexual partners with the 14 nucSNVs significantly associated with the number of sexual partners in males and all 265 mtSNVs. Again, we found no interaction term to be significantly different from 0 after FDR correction for 3,318 tests. In conclusion, we did not detect any nucSNV with additive effects to compensate for the 265 mtSNVs’ effects on the number of offspring or the number of sexual partners in males, but we cannot exclude the possibility that some other nucSNVs do. In particular, future studies focusing on nucSNVs in genes impacting mitochondrion functions may be productive, although detecting epistasis generally suffers from a low statistical power because of the huge number of possible interacting SNV pairs. Additionally, nuclear compensatory mutations for MC variants are rapidly fixed because they are positively selected in males (Osada and Akashi 2012; Connallon et al. 2018). Hence, there may be very few unfixed compensatory nucSNVs in human populations.

We used the number of offspring of an individual as a measure of this person's realized fitness. It should be noted that, in modern societies, this measure may differ substantially from fertility because of many biological and nonbiological factors. Regardless, the realized fitness instead of fertility is what matters to evolution. This said, we caution that a moderate alteration in fertility may affect the number of offspring in premodern societies but not today. Following a previous study (Zietsch et al. 2021), we used the number of sexual partners as an approximate measure of premodern reproductive fitness. However, the number of sexual partners in modern societies may differ from that in premodern societies due to cultural and societal changes. In other words, the number of sexual partners in modern societies may be only moderately correlated with premodern reproductive fitness. Nonetheless, this proxy is the best estimate we have for premodern reproductive fitness.

We shall consider the possibility that MC variants are abundant but are undetected because our study lacks sufficient statistical power. The UKB genotyped only 265 mtSNVs, while there are at least 4,116 mtSNVs according to the human mitochondrial genome phylogeny (van Oven and Kayser 2009), meaning that we used only 6.4% of all mtSNVs. It is possible that mtSNVs conforming to MCH were not genotyped. Notwithstanding, the genotyped mtSNVs in the UKB are relatively common (The UK Biobank 2014); having information for rarer mtSNVs may not be helpful in improving statistical power. Further, although mitochondrial whole-genome sequencing data are available in the “All of Us” (AoU) program (The All of Us Research Program Investigators 2019), AoU does not contain data of the number of offsring. In fact, our study has at least some statistical power, but the findings do not support MCH. For example, we detected a significant positive instead of negative intersexual correlation in the effect on the number of offspring across mitochondrial haplogroups. We detected a significant sex-haplogroup interaction in the effect on the number of sexual partners, but the interaction is opposite in direction to the MCH prediction. We also discovered sex-differential survival effects of mtDNA variants, but the pattern does not support MCH. Finally, we found sex-specific disease associations of mtSNVs, but the parity between sexes in the number of associations does not support MCH. These findings suggest that our conclusion will likely remain the same, at least qualitatively, even when a larger sample and a higher statistical power are acquired. Therefore, despite the previous report of a weak MC variant (Milot et al. 2017), mother's curse does not play a prominent role in human mitochondrial genome evolution and, based on data from the UK Biobank, the overall mtDNA mutation load is similar in the two sexes.

Materials and Methods

Data Source

All genomic and phenotypic data in this study were obtained from the UKB (Bycroft et al. 2018) under project no. 882999. The phenotypic data included the number of children (data-field 2734 for females and data-field 2405 for males), the number of opposite-sex sexual partners (data-field 3669, data-field 2139, and data-field 2149), age at recruitment (data-field 21022), year of birth (data-field 34), and diseases (data-field 41270 and data-field 20002). The number of opposite-sex sexual partners was calculated by subtracting the number of same-sex sexual partners from the total number of sexual partners. Negative values of the number of opposite-sex sexual partners were treated as 0. Only the 19 diseases previously reported to be associated with mtDNA variants in mtGWAS with combined sexes (Yonova-Doing et al. 2021) were considered in this study. The genomic data used included whole-genome sequencing data of the nuclear genome and mitochondrial genotyping data. The first ten PCs of the nuclear genome were obtained from the UKB.

Data Rescaling

For the number of opposite-sex sexual partners, values greater than 100 were treated as 100 because extreme outliers can strongly bias the results. For the number of opposite-sex sexual partners ranging from 0 to 100, because the raw data are highly right-skewed even after being capped at 100 (skewness = 5.18), we transformed the raw number of sexual partners (N_raw) to the ln-transformed number of sexual partners (N_transformed) by N_transformed = ln(N_raw + 1). While the cutoff of 100 was arbitrarily chosen, we found no qualitative difference in the results of the linear models, PGLS models, and interaction models when we changed the cutoff to 50, 150, or 200. With any chosen cutoff, no haplogroup has a significant effect on either sex except for haplogroup H2a2b5a's effect on males (adjusted P = 0.082 at cutoff 50; 0.051 at cutoff 100; 0.045 at cutoff 150; and 0.044 at cutoff 200). Furthermore, male HEPN does not significantly predict female HEPN in the PGLS model, and the sex × H2a2b5a interaction is the only significant interaction term in the interaction model.

Data rescaling and statistical analyses were conducted in R (version 4.4.2).

Mitochondrial Genome-wide Association Study

mtGWAS was conducted in PLINK 1.9 (Chang et al. 2015) and PLINK 2 (Chen et al. 2019). Quality control of mtGWAS followed a previous study (Yonova-Doing et al. 2021). The UKB classified participants into European or non-European ancestry based on their nuclear genetic PCs and generated a list of participants with European ancestry. Samples with European ancestry were kept according to the list. The UKB calculated relatedness among participants and output pairs of participants related up to the third degree. For each pair of related individuals, we removed the individual with the higher sample missing rate of mitochondrial genotyping data. 387,310 samples (208,882 females and 178,428 males) and 265 mtSNVs passed all filters. Both the haplogroup-based association analysis and mtGWAS used quality-controlled data from the above procedures. mtGWAS for the number of offspring and number of opposite-sex sexual partners was performed with the covariates being the first ten nuclear genetic PCs and year of birth. mtGWAS for diseases was performed with the covariates being the first ten nuclear genetic PCs and age, using logistic regressions. mtGWAS was conducted separately for males, females, and the two sexes combined.

Nuclear Genome-wide Association Study

Quality control of nuclear genome-wide association study (nucGWAS) followed a previous study (Song and Zhang 2024). Variants with minor allele frequencies above 0.01 passed the filter. Variants with genotype missing rates below 0.02 passed the filter. Variants violating Hardy–Weinberg equilibrium with P-values smaller than 10⁻¹⁰ were removed. Imputed variants with INFO scores above 0.8 passed the filter. Linkage disequilibrium (LD) pruning was performed with the following procedure. First, to measure LD, we computed r² for all single-nucleotide polymorphism (SNP) pairs in a window of 50 SNPs. For each pair with r² exceeded 0.2, the SNP with the lower minor allele frequency was removed. When no pair of SNPs had an r² above 0.2, the algorithm moved five SNPs forward to the next window and repeated the process. The above LD-pruning was conducted with parameter “–indep-pairwise 50 5 0.2” in PLINK 2. Samples with sample missing rates below 0.02 passed the filter. Samples with non-European ancestry were removed following the procedure in mtGWAS. Samples with inbreeding coefficients beyond the range of mean ± 3 standard deviations were removed. Up to three-degree relatives were removed following the procedure in mtGWAS. In total, 369,103 samples (198,884 females, 170,219 males) and 978,960 variants passed all filters. nucGWAS for the number of offspring and nucGWAS for the number of opposite-sex sexual partners were performed for males, with the covariates being the first ten nuclear genetic PCs and year of birth.

Testing Epistasis Between mtSNVs and nucSNVs

A total of 155,441 samples passed both mtGWAS and nucGWAS quality controls and were used in the test of nuclear compensation for mtSNVs. Note that the analysis was performed for males only, because only the nuclear compensation for MC variants in males is relevant to MCH. nucSNVs with nominal P-values below 5 × 10⁻⁸ were considered significantly associated with the trait concerned. All 265 mtSNVs and the 12 nucSNVs significantly associated with the number of offspring in males were included in the following compensation analysis for the number of offspring. For each pair of mtSNV and nucSNV, an interaction model was built to test the interactive effect of the mtSNV and the nucSNV on the number of offspring. The model predicted the number of offspring by ten nuclear genetic PCs, year of birth, the mtSNV, the nucSNV, and the interaction between the mtSNV and the nucSNV. The P-values of the interaction terms from all models were corrected by FDR. An adjusted P-value below 0.05 was considered a significant mtSNV–nucSNV interaction. A similar analysis was conducted for the trait of the number of opposite-sex sexual partners with all 265 mtSNVs and the 14 nucSNVs significantly associated with the number of opposite-sex sexual partners in males.

Mitochondrial Haplogroup Classification

The Revised Cambridge Reference Sequence (rCRS) (Andrews et al. 1999) was used as the mitochondrial genome reference sequence. Individuals were classified into mitochondrial haplogroups by Haplogrep 3 (Schönherr et al. 2023) based on the mitochondrial genotyping data. rCRS tree version 17.2 was used in haplogroup classification. Haplogrep 3 generates a quality score for each individual that measures the confidence of classifying the individual into the haplogroup (Fig. S2). In the haplogroup-based association analysis, we removed individuals with a quality score below 0.8 after the above quality control procedure in the mtGWAS section. A total of 378,777 samples (204,240 females and 174,537 males) passed the filter.

Linear Models for Predicting the Number of Offspring

The OLS model was used to predict the number of offspring by the year of birth, mitochondrial haplogroup, and ten nuclear genetic PCs. Note that the year of birth, ranging from 1934 to 1970 for the participants in our data, was used as a covariate to control the birth cohort rather than the age, because census data showed that the number of offspring decreased nearly linearly with the year of birth (Office for National Statistics 2025). Similarly, the year of birth was used as a covariate in the model predicting the number of sexual partners because the number of sexual partners tended to increase with the year of birth due to changes in social norms (Mercer et al. 2013). Males and females have separate models. For the ith participant in the model, let

$$I(i,k)=\{\begin{array}{c}1,\mathrm{if}i\mathrm{is}\mathrm{of}\mathrm{haplogroup}k,\\ 0,\mathrm{if}i\mathrm{is}\mathrm{not}\mathrm{of}\mathrm{haplogroup}k.\end{array}$$

Then, the number of offspring of the ith participant (N_i) was predicted by

$$N_i={\beta }_0+{\beta }_{1}YO{B}_i+{\displaystyle \sum _{j=1}^{10}}{\beta }_{j+1}P{C}_{j,i}+{\displaystyle \sum _{k=1}^{602}}{\beta }_{k+11}I(i,k)+{\epsilon }_i$$ Model 1

Here, ${\beta }_x$ is the regression coefficient of the xth predictor variable, YOB is year of birth, $P{C}_{j,i}$ is the jth PC of the ith observation in the first ten PCs, and $\epsilon$ is the error term.

The interaction model adds interaction terms between sex and each of the other predictor variables to Model 1. For the ith participant in the model, let

$$SE{X}_i=\{\begin{array}{c}1,ifiismale\\ 0,ifiisfemale\end{array}$$

Then, the number of offspring of the ith participant (N_i) was predicted by

$$\begin{array}{rl}{N}_i& ={\beta }_0+{\beta }_{1}YO{B}_i+{\beta }_{2}SE{X}_i+{\beta }_{3}SE{X}_{i}YO{B}_i\\ & +{\displaystyle \sum _{j=1}^{10}}({\beta }_{j+3}P{C}_{j,i}+{\beta }_{j+13}P{C}_{j,i}SE{X}_i)+\end{array}$$

$$\begin{array}{rl}{N}_i& ={\beta }_0+{\beta }_{1}YO{B}_i+{\beta }_2{\mathit{S}\mathit{E}\mathit{X}}_i+{\beta }_3{\mathit{S}\mathit{E}\mathit{X}}_{i}YO{B}_i\\ & +{\displaystyle \sum _{j=1}^{10}}({\beta }_{j+3}P{C}_{j,i}+{\beta }_{j+13}P{C}_{j,i}{\mathit{S}\mathit{E}\mathit{X}}_i)\\ & +{\displaystyle \sum _{k=1}^{602}}({\beta }_{k+23}I(i,k)+{\beta }_{k+625}I(i,k){\mathit{S}\mathit{E}\mathit{X}}_i)+{\epsilon }_i\end{array}$$ Model 2

All variables are defined as in Model 1. $I(i,k)SE{X}_i$ is the sex × haplogroup interaction term.

A total of 603 haplogroups had at least one sample per sex for the number of offspring and were included in model fitting. Haplogroup H1 was chosen as the reference haplogroup in the linear model because (i) H1 is the haplogroup with the largest sample size and (ii) the mean number of offspring of individuals of H1 is not significantly different from the mean of all other haplogroups combined in either sex (P = 0.56 for males and 0.70 for females, two-tailed t-test; Table S1). The reference was needed to avoid redundant parameters. In Model 1, the regression coefficient of a haplogroup is the effect of the haplogroup minus the effect of H1 on the number of offspring in the sex concerned. In Model 2, the regression coefficient of a sex × haplogroup interaction term is the effect of the haplogroup in males minus the effect of the haplogroup in females on the number of offspring.

Phylogenetic Generalized Least Squares Models

The phylogenetic tree of human mitochondrial haplogroups was obtained from https://phylotree.org/(van Oven and Kayser 2009) and converted to the Newick format. The number of mutations accumulated was used to measure branch lengths of the tree. A PGLS model was built by the R function pgls() in the caper package (Freckleton et al. 2002) based on the haplogroup effect on the number of offspring estimated in preceding sections. The effect in females was the response variable, while the effect in males was the predictor variable. In the PGLS model, three parameters were used to transform phylogenetic covariances and branch lengths to reflect different assumptions: phylogenetic signal strength (λ), graduality of trait evolution along branches (κ), and timing in divergence (δ). To keep the model simple, we fixed all three parameters at 1, meaning that the trait covariance between lineages exactly matches the Brownian motion model's prediction and all branch lengths are on the original scale. A similar PGLS analysis was performed for the number of opposite-sex sexual partners. The results of PGLS models were robust to the parameter choice because the significance or lack of significance did not change when we estimated all parameters by maximum likelihood and used these estimated parameters instead of 1.

mtDNA Variants With Different Frequencies in the Two Sexes

For each biallelic mtSNV, a 2 × 2 contingency table was used to test whether the allele frequency differs between the two sexes. mtSNVs with fewer than six counts in any cell were filtered out due to the sample size requirement of chi-squared tests. A chi-squared test was then performed, followed by an FDR correction. mtSNVs with adjusted P < 0.05 were considered to have significantly different frequencies in the two sexes. We similarly analyzed each mitochondrial haplogroup, where all other haplogroups combined were considered as the alternative allele.

mtGWAS vs. Haplogroup-Based Association Analysis

We used two approaches to study the association between mtDNA variants and a phenotypic trait: mtSNV-based mtGWAS and haplogroup-based association analysis. The two approaches have their respective pros and cons. mtGWAS is expected to have a higher statistical power for the genotyping data used because there are fewer mtSNVs than haplogroups. However, a commonly used assumption underlying mtGWAS is that mtSNVs have independent, additive effects on the trait of concern. Because the human mitochondrial genome is in complete linkage, the above assumption is violated. In addition, mtSNVs identified by mtGWAS are less informative than haplogroups. For instance, A11251G, C15452A, and T4216C are all associated with ptosis of eyelid (Table 1), which can be misinterpreted as three independent mtSNVs associated with the disease. In fact, these three mtSNVs define the haplogroup R2'JT. Furthermore, in contrast to nuclear GWAS, the complete linkage of mtDNA means that fine mapping for identifying causal variants is infeasible; it is possible that the causal mtSNV is distant from the mtSNV found to be associated with the trait.

By contrast, haplogroup-based association analysis accounts for epistasis and linkage. However, the genotyping data only allow classifying mtDNA genotypes into known haplogroups. When a mutation creates a new haplogroup that is not in the reference panel, the mtDNA carrying the mutation will be filtered out or classified as the most similar haplogroup in the reference panel. As a result, MC variants not in the reference panel would be missed. Another issue of haplogroup-based association analysis is that some haplogroups are too similar (e.g. differing by potentially functionally irrelevant mutations). This fine classification could lower the statistical power, especially if related haplogroups share a causal mutation. To tackle this issue, some studies classify mtDNAs into macro-haplogroups, which are phylogenetic clusters of haplogroups (Yonova-Doing et al. 2021). However, there is little empirical evidence that using macro-haplogroups is statistically more efficient than using haplogroups.

Supplementary Material

Supplementary material is available at Genome Biology and Evolution online.

Data Availability

Relevant data and code can be obtained from Zenodo (https://doi.org/10.5281/zenodo.17576174).